Project Axiom
The Silicon-Native, Zero-Overhead Multimodal AI Engine engineered from scratch in modern C++23 & bare-metal NVIDIA CUDA.
Overview
Project Axiom is a hardcore, bare-metal rewrite of deep learning mechanics engineered directly in C++23 and NVIDIA CUDA. It features true native multimodality via a discrete VQ-VAE integer space, zero-latency pinned memory PCIe DMA, asynchronous ping-pong buffering, and CUTLASS v4.8 Tensor Core saturation.
Problem, Approach, Outcome
Stripping away framework bloat, interpreter overhead, and VRAM fragmentation by implementing custom device kernels, dual-precision master weights, and CUTLASS Epilogue Visitor Trees (EVT).
“An idiot admires complexity; a genius admires simplicity.”
Phase 2 scaling includes BFC caching allocators, PagedAttention for KV cache management, CUDA Graphs for zero-launch overhead, and dedicated desktop/web UI interfaces.